Tensor parallelism is all you need. Run LLMs on weak devices or make powerful devices even more powerful by distributing the workload and dividing the RAM usage.
Why do you think that https://github.com/beam-cloud/beta9 is a good alternative to distributed-llama